Retrieval, agents,and interfaces.

B.Tech in Computer Science and Engineering (Data Science) at Sreyas Institute of Engineering and Technology, Hyderabad. October 2022 to June 2026.

The certifications covered the three layers I wanted to work in: Core Python, front-end development, and generative AI on Oracle Cloud Infrastructure.

• Python
• TypeScript
• React
• FastAPI
• PostgreSQL / pgvector
• Qdrant
• LangGraph
• LangSmith

A stock research and trading platform. My work covered a table of 1,000+ ticking rows, a streaming news page, observability, and experiments with OCR and vision-language models.

• 01Live tableRows updated continuously. I virtualized the table so only visible rows mounted: 40% faster first paint and 60% less memory.
• 02MarketBuzzNews page that streams live backend data and merges RSS feeds into the same view.
• 03ObservabilityStructured logs, metrics and Amazon CloudWatch alarms on the services I worked on, so failures had logs behind them.
• 04Models and visionExperimented with Claude's open SDKs, OCR models and vision-language models. Evaluated open-source Hugging Face models, including Chinese models, for cost, accuracy and scale.

The constraint was memory and paint time: every mounted row costs something while the data keeps moving.

At Tradl AI, I worked on the RAG pipeline behind the platform's stock research answers. The documents were financial reports, and the path ran from ingestion and cleaning to hybrid retrieval, ranking and evaluation. When retrieval missed the right page, the model had nothing useful to answer from.

1. 01Extraction
2. 02Cleaning
3. 03Indexing
4. 04Hybrid retrieval
5. 05Ranking
6. 06Answer

• ExtractionDocling and MinerU for clean extraction from financial reports, then data cleaners before anything reached the index.
• End to endHandled the RAG path over financial reports: extraction, cleaning, indexing, hybrid retrieval, ranking and the benchmark.
• Vision experimentsRAG experiments with OCR and vision-language models for document extraction and retrieval.
• Benchmark90 questions. Recall measured on the retrieval layer, separately from the model, before and after the rebuild.
• Result0.37 to 0.61 on the same 90 questions: plus 0.24 absolute, about 1.65 times the recall.

The failure is silent: a missing page shows up as a worse answer, not an error. Extraction quality sets the ceiling, so cleaning came before ranking.

Token compression, page-level RAG, multi-agent research, arXiv summaries, an agent dashboard, and text-to-query. Each one targets a specific failure in AI products: cost, citations, confidence or visibility.

Each has a failure path: a test suite, a critic, an ambiguity gate, a human checkpoint or a live event feed.